In TensorRT 10.x, INT8 conversion requires FP16 to be set as a fallback for
network layers (e.g. normalization ops in detection models) that have no INT8
kernel implementation. Without FP16, build_serialized_network can return None
on Jetson for YOLO-type models. INT8 flag is still the primary precision;
FP16 is only the layer-level fallback within the same engine.
Made-with: Cursor
- Dockerfile.jetson: JetPack 6.x L4T base image (aarch64), TensorRT and PyCUDA from apt
- requirements-jetson.txt: derived from requirements.txt, no pip tensorrt/pycuda
- docker-compose.jetson.yml: runtime: nvidia for NVIDIA Container Runtime
- tensorrt_engine.pyx: convert_from_source accepts optional calib_cache_path; INT8 used when cache present, FP16 fallback; get_engine_filename encodes precision suffix to avoid engine cache confusion
- inference.pyx: init_ai tries INT8 engine then FP16 on lookup; downloads calibration cache before conversion thread; passes cache path through to convert_from_source
- constants_inf: add INT8_CALIB_CACHE_FILE constant
- Unit tests for AC-3 (INT8 flag set when cache provided) and AC-4 (FP16 when no cache)
Made-with: Cursor
Oleksandr Bezdieniezhnykh